xref: /aosp_15_r20/external/XNNPACK/src/qs8-igemm/gen/2x4c2s4-minmax-fp32-wasmsimd-dot16x2-ld128.c (revision 4bdc94577ba0e567308109d787f7fec7b531ce36)
1 // Auto-generated file. Do not edit!
2 //   Template: src/qs8-igemm/MRx4c2s4-wasmsimd-dot16x2.c.in
3 //   Generator: tools/xngen
4 //
5 // Copyright 2021 Google LLC
6 //
7 // This source code is licensed under the BSD-style license found in the
8 // LICENSE file in the root directory of this source tree.
9 
10 #include <assert.h>
11 
12 #include <wasm_simd128.h>
13 
14 #include <xnnpack/igemm.h>
15 #include <xnnpack/math.h>
16 
17 
xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128(size_t mr,size_t nc,size_t kc,size_t ks,const int8_t ** restrict a,const void * restrict w,int8_t * restrict c,size_t cm_stride,size_t cn_stride,size_t a_offset,const int8_t * zero,const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS (1)])18 void xnn_qs8_igemm_minmax_fp32_ukernel_2x4c2s4__wasmsimd_dot16x2_ld128(
19     size_t mr,
20     size_t nc,
21     size_t kc,
22     size_t ks,
23     const int8_t** restrict a,
24     const void* restrict w,
25     int8_t* restrict c,
26     size_t cm_stride,
27     size_t cn_stride,
28     size_t a_offset,
29     const int8_t* zero,
30     const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS
31 {
32   assert(mr != 0);
33   assert(mr <= 2);
34   assert(nc != 0);
35   assert(kc != 0);
36   assert(ks != 0);
37   assert(ks % (2 * sizeof(void*)) == 0);
38   assert(a_offset % sizeof(int8_t) == 0);
39   assert(a != NULL);
40   assert(w != NULL);
41   assert(c != NULL);
42 
43   int8_t* c0 = c;
44   int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
45   if XNN_UNPREDICTABLE(mr != 2) {
46     c1 = c0;
47   }
48 
49   kc = round_up_po2(kc, 8 * sizeof(int8_t));
50   do {
51     v128_t vacc0x0123 = wasm_v128_load(w);
52     v128_t vacc1x0123 = vacc0x0123;
53     w = (const void*) ((const int32_t*) w + 4);
54 
55     size_t p = ks;
56     do {
57       const int8_t* restrict a0 = a[0];
58       if XNN_UNPREDICTABLE(a0 != zero) {
59         a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
60       }
61       const int8_t* restrict a1 = a[1];
62       if XNN_UNPREDICTABLE(a1 != zero) {
63         a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
64       }
65       a += 2;
66 
67       size_t k = kc;
68       do {
69         v128_t vxa0 = wasm_i16x8_load8x8(a0);
70         a0 += 8;
71         v128_t vxa1 = wasm_i16x8_load8x8(a1);
72         a1 += 8;
73 
74         const v128_t vb01 = wasm_v128_load(w);
75         const v128_t vxb0 = wasm_i16x8_extend_low_i8x16(vb01);
76         const v128_t vxb1 = wasm_i16x8_extend_high_i8x16(vb01);
77 
78         vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb0));
79         vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
80         vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb0));
81         vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
82 
83         vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb1));
84         vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
85         vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb1));
86         vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
87         const v128_t vb23 = wasm_v128_load((const int8_t*) w + 16);
88         const v128_t vxb2 = wasm_i16x8_extend_low_i8x16(vb23);
89         const v128_t vxb3 = wasm_i16x8_extend_high_i8x16(vb23);
90 
91         vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb2));
92         vxa0 = wasm_v32x4_shuffle(vxa0, vxa0, 1, 2, 3, 4);
93         vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb2));
94         vxa1 = wasm_v32x4_shuffle(vxa1, vxa1, 1, 2, 3, 4);
95 
96         vacc0x0123 = wasm_i32x4_add(vacc0x0123, wasm_i32x4_dot_i16x8(vxa0, vxb3));
97         vacc1x0123 = wasm_i32x4_add(vacc1x0123, wasm_i32x4_dot_i16x8(vxa1, vxb3));
98 
99         w = (const int8_t*) w + 32;
100         k -= 8 * sizeof(int8_t);
101       } while (k != 0);
102       p -= 2 * sizeof(void*);
103     } while (p != 0);
104 
105     vacc0x0123 = wasm_f32x4_convert_i32x4(vacc0x0123);
106     vacc1x0123 = wasm_f32x4_convert_i32x4(vacc1x0123);
107 
108     const v128_t vscale = wasm_v128_load64_splat(params->fp32_wasmsimd.scale);
109     vacc0x0123 = wasm_f32x4_mul(vacc0x0123, vscale);
110     vacc1x0123 = wasm_f32x4_mul(vacc1x0123, vscale);
111 
112     const v128_t vmagic_bias = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias);
113     vacc0x0123 = wasm_f32x4_add(vacc0x0123, vmagic_bias);
114     vacc1x0123 = wasm_f32x4_add(vacc1x0123, vmagic_bias);
115 
116     const v128_t vmagic_min = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_min);
117     vacc0x0123 = wasm_i32x4_max(vacc0x0123, vmagic_min);
118     vacc1x0123 = wasm_i32x4_max(vacc1x0123, vmagic_min);
119 
120     const v128_t vmagic_bias_less_output_zero_point = wasm_v128_load64_splat(params->fp32_wasmsimd.magic_bias_less_output_zero_point);
121     vacc0x0123 = wasm_i32x4_sub(vacc0x0123, vmagic_bias_less_output_zero_point);
122     vacc1x0123 = wasm_i32x4_sub(vacc1x0123, vmagic_bias_less_output_zero_point);
123 
124     v128_t vacc01x0123 = wasm_i16x8_narrow_i32x4(vacc0x0123, vacc1x0123);
125 
126     v128_t vout = wasm_i8x16_narrow_i16x8(vacc01x0123, vacc01x0123);
127 
128     const v128_t voutput_max = wasm_v128_load64_splat(params->fp32_wasmsimd.output_max);
129     vout = wasm_i8x16_min(vout, voutput_max);
130 
131     if (nc >= 4) {
132       *((float*) c1) = (float) wasm_f32x4_extract_lane(vout, 1);
133       *((float*) c0) = (float) wasm_f32x4_extract_lane(vout, 0);
134 
135       c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
136       c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
137 
138       a = (const int8_t**restrict) ((uintptr_t) a - ks);
139 
140       nc -= 4;
141     } else {
142       uint32_t vout1 = wasm_i32x4_extract_lane(vout, 1);
143       uint32_t vout0 = wasm_i32x4_extract_lane(vout, 0);
144       if (nc & 2) {
145         *((uint16_t*) c1) = (uint16_t) vout1;
146         vout1 >>= 16;
147         c1 += 2;
148         *((uint16_t*) c0) = (uint16_t) vout0;
149         vout0 >>= 16;
150         c0 += 2;
151       }
152       if (nc & 1) {
153         *c1 = (int8_t) vout1;
154         *c0 = (int8_t) vout0;
155       }
156 
157       nc = 0;
158     }
159   } while (nc != 0);
160 }
161